MirrorCode

Adoption judgment: MirrorCode is worth considering as a high-attention benchmark for long-horizon AI software engineering, not as a development product that teams can simply purchase and use. The available evidence supports the value of its evaluation setup, but does not show that leaderboard scores equal real-world productivity or that a model will perform reliably across all engineering work.

What it does: MirrorCode treats the target software as a black box. Agents cannot access the original source code or the internet; they must infer behavior from the executable environment and available feedback, rebuild the project from scratch, and pass both visible and hidden tests. Posts from Epoch AI Research and the project author describe the sandbox, anti-cheating constraints, and multi-day autonomous coding setup. The cited leaderboard reports a 64% solve rate for Claude Fable 5 and 20% for GPT-5.6 Sol. These are results for a particular benchmark configuration, not a general ranking of programming ability.

Barrier and cost: This is not an out-of-the-box coding utility.